Clustering of Visual Data using Antinspired Methods

1
Clustering of Visual Data using Ant-inspired
Methods
Tijana Janjusevic Multimedia and Vision
Group, Queen Mary, University of London
Supervisor Prof. Ebroul Izquierdo
2
Overview

• Introduction
• Biologically Inspired Optimization Systems
• ACO based image classifier
• Ant Colony Optimization (ACO)
• Subspace clustering by Ants
• Experimental Results
• Ant-tree for video summarization
• Ant-tree - new model for clustering
• Experimental Results
• Future work

3
Image Classification

• Data Mining tasks feature extraction,
• pattern
  recognition,
• segmentation,
• feature selection,
• classification,
• 
  optimization,
• annotation,
  
• Image Classification
• the task is to learn to assign images with same
  semantic content to predefined classes
• two types of classification schemes supervised
  and unsupervised.
• Supervised classification
• requires relevance feed-back from a human
  annotator and training data
• Unsupervised classification - Clustering
  
• without training or need for knowledge about the
  data
• The performance of the image classification
  algorithms relies on the efficient optimisation
  techniques

4
Biologically Inspired Optimisation techniques

• Recent developments in applied and heuristic
  optimisation have been
• strongly influenced and inspired by natural
  and biological system.
• Biologically Inspired systems

Artificial Immune Systems,
Particle Swarm,
Ant Colony Systems
5
Ant Colony Optimisation (ACO)

• Meta-heuristic that uses strategies of ants to
  solve optimization problems.
• An important and interesting behavior of ant
  colonies is their foraging behavior, and, in
  particular, how ants can find shortest paths
  between food sources and their nest, using
  pheromone driven communication.

6
Ant Colony System

• The Ant System algorithm (AS) was first proposed
  to solving the Traveling Salesman Problem (TSP).
• Given a set of n points and a set of distances
  between them, we call dij the length of the
• path between points i and j.
• The probability of choosing next j node
• 
  ? Heuristic
  information
• Pheromone value

7
Problem Definition

• Combination of low-level visual features in
  Clustering
• Each group of images may correlate with respect
  to different set of important features,
• and each group may contain some irrelevant
  features

CLD, EHD
TGF, EHD
CSD, TGF
CSD, CLD
CSD, CLD
CSD, DCD, GLC
Lion Building Rural
Car Elephant
Clouds
8
Subspace Clustering using ACO

• Ant Colony Optimisation and its learning
  mechanism is implemented for optimizing feature
  weights for each cluster of images.

CLD, EHD
TGF, EHD
CSD, TGF

• Each ant clusters images according to
• different local
• feature weights
• pheromone value
• from previous
• solutions

CSD, CLD
CSD, DCD, GLC
CSD, CLD
9
Subspace Clustering using ACO

• Multi-feature Combination
• Proposed Algorithm
• Each ant will assign each image ,
  to the cluster , , with the
  probability obtained from
• Computation of weights
• Computation of centroids.
• Pheromone Update

10
Experimental Evaluation

• Low-level features (descriptors) used For visual
  representation of images
• Synthetic Data
  Results

• Colour Layout (CLD),
• Colour Structure (CSD),
• Dominant Colour (DCD),
• Edge Histogram (EHD)
• Grey Level Co-occurrence Matrix (GLC)

11
Experimental results

• The Corel image database - 600 images with 6
  semantic concepts

Lion Building Rural
Car Elephant Clouds
12
Experimental results

• Flickr Image Database - 500 images segmented into
  regions.
• Semantic Concepts Sand, Sea, Vegetation,
  Building, Sky, Person, Rock, Tree, Grass, Ground,
  and.

13
Ant-tree for video summarisation

• Ant-Tree clustering method
• Inspired by self-assembling behavior of African
  ants and their ability to build chains (bridges)
  by their bodies in order to link leaves together.
• We model the ability of ants to build live
  structures with their bodies in order to
  discover, in a distributed and unsupervised way,
  a tree-structured organisation and summarisation
  of the video data.

14
AntTree New model for clustering

• General principles
• each ant represents node of tree (data)
• - outgoing link ai can maintain
  toward another ant
• - incoming links other ants maintain
  toward ai
• - ao support, apos position of
  moving ant

Support case 1. If no ant is connected
yet to support a0 Then connect ai to a0 2.
Else let a be the ant connected to a0,
which is the most similar to ai
If Sim(ai,a)gtTsim (ai) Then move ai toward
a Else If
Sim(ai,a)ltTdissim(ai) Then connect ai to the
support Else decrease Tsim (ai)
and increase Tdissim(ai)
Tsim(ai) Tsim(ai)0.9
Tdissim(ai) Tdissim(ai)0.01
Ant case 1. let apos denote the ant on
which ai is located and let ak denote a
randomly selected neighbour of apos 2.
If Sim(ai,a)gtTsim (ai) Then connect ai to apos
Else decrease Tdissim(ai), increase Tsim
(ai) and move ai toward ak 3. Else
randomly move ai toward ak

• Main algorithm
• 1. all ants placed on the support
• initialization Tsim(ai)1, Tdissim(ai)0
• 2. While there exists non connected ant ai Do
  
• 3. If ai is located on the support Then
  Support case
• 4. Else Ant case
• 5. End While

15
Surveillance video

• each cluster is represented by support
  (representative) frame
• to make a summary of whole video, one video
  segment is taken from each cluster starting with
  support frame
• to determine the length of video segment the
  maximum number of negative clustered frames going
  consecutively is used as threshold
• maximum length of video segment is 10sec.

16
Summarization results (PETS2001)
Camera video 4 min 30 sec (duration)
Video summary 53 sec
17
Summarization results (gate2)
Gate_2 video 30 min (duration)
Video summary 1 min 16 sec
18
Future work

• Improvement of SC Method
• Pheromone driven mechanism of ACO will be
  used for optimization of clustering
• task by searching for number of clusters that
  leads to best clustering.
• Implementation of FS for Video Summarization and
  Scene Detection
• The aim of this task is to detect and classify
  events from video using
• intelligent combination of multiple low-level
  visual features.

19
Thank you for your attention!
tomas.piatrik_at_elec.qmul.ac.uk